@@ -19,9 +19,9 @@ export default function Welcome() {
19
19
{ feature : { atom : 25 * GPT4_ATOMS_PER_SHARD + 8 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_latelayer_postmlp' ] . get_ae ( { } ) } ,
20
20
label : "police reports, especially child safety" , description : "safety incidents especially related to children" } ,
21
21
{ feature : { atom : 9 * GPT4_ATOMS_PER_SHARD + 44 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_latelayer_postmlp' ] . get_ae ( { } ) } ,
22
- label : "price increases " , description : "ends of phrases describing commodity/equity price increases " } ,
22
+ label : "price changes " , description : "ends of phrases describing commodity/equity price changes " } ,
23
23
{ feature : { atom : 17 * GPT4_ATOMS_PER_SHARD + 33 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_latelayer_postmlp' ] . get_ae ( { } ) } ,
24
- label : "ratification (multilingual)" , description : "ends of phrases describing commodity/equity price increases " } ,
24
+ label : "ratification (multilingual)" , description : "ratification (multilingual) " } ,
25
25
{ feature : { atom : 3 * GPT4_ATOMS_PER_SHARD + 421 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_latelayer_postmlp' ] . get_ae ( { } ) } ,
26
26
label : "would [...]" , description : "conditionals (things that would be true)" } ,
27
27
{ feature : { atom : 63 * GPT4_ATOMS_PER_SHARD + 8 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_latelayer_postmlp' ] . get_ae ( { } ) } ,
@@ -56,6 +56,9 @@ export default function Welcome() {
56
56
{ feature : { atom : 733 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_l8_postmlp' ] . get_ae ( {
57
57
num_features : '32768' , num_active_features : '8'
58
58
} ) } , label : "Patrick/Patty surname predictor" , description : "Predicts surnames after Patrick" } ,
59
+ { feature : { atom : 64464 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_l8_postmlp' ] . get_ae ( {
60
+ num_features : '131072' , num_active_features : '32'
61
+ } ) } , label : "things that are unknown" , description : "things that are unknown" } ,
59
62
{ feature : { atom : 56907 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_l8_postmlp' ] . get_ae ( { // similar to 33248
60
63
num_features : '131072' , num_active_features : '32'
61
64
} ) } , label : "words in quotes" , description : "predicts words in quotes" } ,
@@ -80,7 +83,13 @@ export default function Welcome() {
80
83
{ heading : 'Safety relevant features (found via attribution methods)' , heading_type : 'h3' , feature : null , label : '' } ,
81
84
{ feature : { atom : 64840 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_l8_postmlp' ] . get_ae ( {
82
85
num_features : '131072' , num_active_features : '32'
83
- } ) } , label : "profanity" , description : "activates in order to output profanity" } ,
86
+ } ) } , label : "profanity (1)" , description : "activates in order to output profanity" } ,
87
+ { feature : { atom : 104813 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_l8_postmlp' ] . get_ae ( {
88
+ num_features : '131072' , num_active_features : '32'
89
+ } ) } , label : "profanity (2)" , description : "activates on profanity" } ,
90
+ { feature : { atom : 101090 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_l8_postmlp' ] . get_ae ( {
91
+ num_features : '131072' , num_active_features : '32'
92
+ } ) } , label : "profanity (3)" , description : "activates on 'fucking' (profane, not sexual contexts)" } ,
84
93
{ feature : { atom : 72185 , autoencoder : AUTOENCODER_FAMILIES [ 'v5_l8_postmlp' ] . get_ae ( {
85
94
num_features : '131072' , num_active_features : '32'
86
95
} ) } , label : "erotic content" , description : "erotic content" } ,
@@ -101,7 +110,7 @@ export default function Welcome() {
101
110
102
111
return (
103
112
< div className = "flex flex-col" style = { { 'padding' : '100px' } } >
104
- < h1 className = "text-2xl font-bold mb-4" > Welcome! This is a viewer for sparse autoencoders features trained in < a href = "todo " > this paper</ a > </ h1 >
113
+ < h1 className = "text-2xl font-bold mb-4" > Welcome! This is a viewer for sparse autoencoders features trained in < a href = "https://cdn.openai.com/papers/sparse-autoencoders.pdf " > this paper</ a > </ h1 >
105
114
< h1 > Pick a feature:</ h1 >
106
115
< FeatureSelect
107
116
init_feature = { feature }
0 commit comments